cataloom

Turn any catalog into a conversation.

cataloom is an open-source Generative UI framework for building conversational, agentic-RAG websites. Instead of a search box and a grid of results, visitors describe what they want in plain language — "SUV automática, menos de $25k, poco kilometraje" — and cataloom retrieves matching items from your content and renders live, interactive UI cards (results, comparisons, quick replies, lead capture) directly in the chat.

It is stack-native to Cloudflare Workers and built around a small, pluggable core so that each domain — cars, real estate, professional services, anything with a catalog — ships as its own vertical plugin.

Status: early / bootstrapping. This repo is the foundation. The first vertical plugin (an automotive "AI Car Finder", available as an optional seed) and the core engine are being built out. Contributions and new vertical plugins are welcome — see CONTRIBUTING.md.

Why

Traditional catalog sites bury intent under filters and forms. A conversational, generative interface:

• lifts conversion — visitors state intent naturally and get an answer, not a filter maze;
• grounds every answer in your content (retrieval-augmented) so the UI never invents a price or a spec;
• renders real UI, not just text — cards, comparisons, and calls-to-action are generated per turn from the actual result set.

How it works

User intent (natural language)
        │
        ▼
  Retrieval (Agentic RAG over your catalog content)
        │
        ▼
  Generative UI response  ──►  a validated UI-component message
        │                       (cards · comparisons · quick replies · lead capture)
        ▼
  Rendered live in the chat surface

• Content in, UI out. Your catalog lives as structured content; retrieval grounds the model on it; the model responds with a UI message (a declarative component array), validated before it ever reaches the browser.
• Pluggable verticals. The core knows nothing about cars or houses. A vertical plugin supplies its content schema, retrieval mapping, and the cards it wants rendered.
• Optional seed content. Ship an empty core, or seed a working demo (e.g. the automotive catalog) so newcomers get a running example in one command.

Tech direction

• Edge-native: Cloudflare Workers (frontend + backend), zero-server.
• Retrieval: managed hybrid search over the catalog content.
• Generative UI: a declarative, validated component protocol (the model returns UI, not HTML).
• Frontend: SvelteKit SSR. Backend: Hono + TypeScript.

The concrete architecture and the first milestone plan live in docs/ARCHITECTURE.md.

Roadmap (short)

1. Core engine — content seam, retrieval, generative-UI response + validator, chat surface.
2. First vertical plugin — automotive "AI Car Finder" (optional seed content).
3. Plugin API — stabilize the contract so the community can add verticals (real estate, professional services, …).

License

MIT — © Nimblersoft. Use it, fork it, build on it.